Carlini Uses LLMs to Port Stockfish to a 1996 Pentium Pro: An Estimated 3300 Elo, With 720 of 1,335 Points of Progress Coming From Algorithms

In a November 1 post on his personal blog, researcher Nicholas Carlini took on a question: in 1996, could an ordinary desktop have beaten Kasparov without building Deep Blue? He bought an old 200 MHz Pentium Pro with 64MB of RAM and had Fable 5.1, Opus 5.5 and GPT-6 write the code to turn Stockfish 17 into a single-core, 32-bit, no-SIMD build retuned for that machine. It measured 3180 ± 80 Elo at 30 seconds per move, extrapolated to about 3300 at classical time controls, above Kasparov's rating of just under 2800 at the time. A second experiment brought the 1996 engine Crafty 11.13 forward with modern search and NNUE evaluation. His conclusion: of the roughly 1,335 Elo gained from the best engines on 1996 commodity hardware to today, 720 points come from better programs rather than faster hardware. In a 100-game match, the modern engine on the Pentium Pro scored 78 wins, 17 draws and 5 losses against the 1996 Crafty running on a modern computer.

Did It Take a Deep Blue?

Deep Blue, which beat Kasparov in 1997, used 30 P2SC processors and 480 custom chess co-processors and could analyze more than 200 million positions per second. In a November 1 blog post, researcher Nicholas Carlini asked the question differently: with today's algorithms, would an ordinary desktop of the time have been enough?

He bought a used machine: a 200 MHz Pentium Pro (released in November 1995) with 64MB of RAM. It boots from a 1.44MB floppy into a minimal system whose only job is to start the engine and the network card, and the engine is online for readers to play. He notes that whenever he says "I wrote" the code, he means he asked Fable 5.1, Opus 5.5 and GPT-6 to write it; only the article itself is his own writing.

Two Experiments

  • Taking Stockfish back to 1996: single core, no SIMD, no 64-bit registers and no C++ standard library, with threading and endgame tablebases removed and the 71MB NNUE replaced by a small network. The naive port scored about 2496 at 1 second per move; after shrinking the network and retuning, it reached 2846 ± 35, and 3180 ± 80 at 30 seconds per move, extrapolated to about 3300 at classical time controls.
  • Bringing Crafty 11.13 forward: as shipped in 1996, it rated 1999 ± 90 on this machine at 1 second per move. Adding modern pruning and reductions, SPSA tuning and an NNUE evaluation lifted it to 2718 ± 34.

In a 100-game match, the modern engine on the Pentium Pro scored 78 wins, 17 draws and 5 losses against the 1996 Crafty running on modern hardware.

What It Suggests About Language Models

Carlini estimates that of the roughly 1,335 Elo separating the best engines on 1996 commodity hardware (about 2300) from today's Stockfish (about 3700), 720 points come from better programs. He acknowledges the ratings come from computer rating lists that do not map directly onto human ratings, and that 3300 is an extrapolation. He closes by turning the question to language models: will software advances over the next few years outpace gains from hardware scale? The toy experiment cannot answer that, but it is a reminder that the sense at the time that few ideas were left to find turned out to be wrong.

via: Nicholas Carlini: Could a modern chess engine on a Pentium have defeated Kasparov in 1996?